Safety / symptom
The reporter has experienced repeated whole-host crashes during Code Intel work, including another crash during an attempted fix. Do not reproduce, build, or run the full test suite on the affected Windows host. This issue tracks a verified resource-pressure defect and a separate, unproven crash attribution.
Read-only evidence
- Earlier Windows records include Kernel-Power 41 with BugcheckCode 26 (0x1A), parameter 1 0x3F; Microsoft documents this subtype as a pagefile inpage CRC error. A separate reboot recorded bugcheck code 0. Do not conflate the two or assign either to this process without process/dump evidence.
- A roughly 5.27 MiB minidump exists from the earlier failure, but the current account cannot read it. Process-creation telemetry for the failure window was not available. The newly reported crash has not been inspected. No dump or machine-specific path is attached.
crates/code-intel-cli/src/snapshot.rs: digest_worktree reads each scoped file with fs::read and retains framed contents in records (lines 1013-1099); hash_records concatenates those records into a second canonical buffer (lines 1610-1616). Peak allocation grows with whole-tree content, not just one file. The successful native-lite path builds snapshots repeatedly; this is a static call-path count, not a measured crash-time RSS.
- MCP
get_gate_verdict calls freshness (mcp_serve/handlers.rs:50-54); committed_evidence.rs:88-120 rebuilds the current worktree snapshot per such request. The idle stdio server does not automatically perform this scan.
Upstream-owned repair and acceptance
- In an independent, safe host/worktree, replace whole-tree retained records plus concatenated canonical buffer with incremental hashing of the same framed bytes and ordering. Keep snapshot identity and artifact-contract parity; bound peak memory by the largest file/chunk rather than the sum of all files. Review whether repeated unchanged snapshots can be avoided without weakening freshness semantics.
- Add deterministic digest-parity and bounded-memory checks with representative large files. Run focused tests plus
cargo test --workspace --no-fail-fast and relevant integration gates on safe CI/host, not on the affected computer. Have an upstream reviewer inspect results before release.
- Treat the host crash as unresolved until a trusted, authorized dump analysis or contemporaneous command/process/commit telemetry establishes causation. A 0x1A/0x3F pagefile CRC can involve storage, controller, memory, or driver faults; Code Intel I/O/memory pressure may trigger that path but is not proven to be its sole cause.
No repository or OS changes, workload reproduction, dump upload, or new crash verification were performed for this report.
Safety / symptom
The reporter has experienced repeated whole-host crashes during Code Intel work, including another crash during an attempted fix. Do not reproduce, build, or run the full test suite on the affected Windows host. This issue tracks a verified resource-pressure defect and a separate, unproven crash attribution.
Read-only evidence
crates/code-intel-cli/src/snapshot.rs:digest_worktreereads each scoped file withfs::readand retains framed contents inrecords(lines 1013-1099);hash_recordsconcatenates those records into a secondcanonicalbuffer (lines 1610-1616). Peak allocation grows with whole-tree content, not just one file. The successful native-lite path builds snapshots repeatedly; this is a static call-path count, not a measured crash-time RSS.get_gate_verdictcalls freshness (mcp_serve/handlers.rs:50-54);committed_evidence.rs:88-120rebuilds the current worktree snapshot per such request. The idle stdio server does not automatically perform this scan.Upstream-owned repair and acceptance
cargo test --workspace --no-fail-fastand relevant integration gates on safe CI/host, not on the affected computer. Have an upstream reviewer inspect results before release.No repository or OS changes, workload reproduction, dump upload, or new crash verification were performed for this report.